Embed Size (px)
Citation preview
Developing the ontological model for research and representation of Commemoration Speeches
in Croatia using a graph property database
Benedikt Perak
Faculty of Humanities and Social Sciences, University of Rijeka
Abstract
This paper demonstrates the ontological model of commemoration speeches given in Croatia that was
developed as a part of the project "Framing the Nation and Collective Identity in Croatia: Political
Rituals and the Cultural Memory of Twentieth Century Traumas". The research gathers linguistic,
media and social data about seven commemoration practices connected to the wars of the 20th century
using various methodologies involving audio-visual field recordings, transcription of commemoration
speeches and the creation of a text corpus. This chapter presents digital humanities methods used to
connect various levels of data analysis and digital resources, from natural language processing (NLP)
of Croatian to conceptual enrichment through the processing of conceptual metaphors. These resources
are integrated using the Neo4j database and the Python Py2Neo library for data manipulation of the
ontological model that embeds the linguistic and media research in the context respectively of the
social identity of actors, their interaction, institutional affiliations and cultural models they represent
and express. The value of this ontological model is in fostering an interdisciplinary approach through
the contextualization of data and targeted usage of digital resources.
Key words
Cultural memory, commemoration, textual analysis, monument representation, social ontology,
complex system, graph database, Neo4j
Introduction
This paper shows the methods and resources used for the digitization of commemoration speeches
given in Croatia and the development of an ontological model for the retrieval and analysis of the
complex socio-cultural data within the project “Framing the Nation and Collective Identity in Croatia:
Political Rituals and the Cultural Memory of Twentieth Century Traumas “(FRAMNAT)1. The
FRAMNAT project seeks to develop digital humanities methodologies applicable for research on
cultural memory and socio-cognitive linguistics analysis. The research involves gathering linguistic,
media and social data about commemoration practices connected to the traumatic events and atrocities
of the 20th century wars: Jasenovac, Bleiburg, Brezovica and Jazovka, Srb, Knin and Vukovar. The
paper presents the procedures of audio-visual and textual data gathering as well the digital humanities
methods used to store and enrich various levels of data analysis using the property graph database
model and resources for natural language processing of Croatian, as well as data integration and
information enrichment.
The motivation for the development of digital methods and tools to study state remembrance practices
is to institute an empirical method of cultural memory research as well as to offer resources and
methods to other researchers of cultural memory in the region and beyond. The systematic analysis,
from fieldwork at the sites of memory to studying the social networks and the role of media in
transmitting narratives, provides an insight into the construction of a society’s “collective memory.”
Although commemorations and political speeches are not the only communication strategies used in
constructing a story of the recent past, commemorations provide a highly visible stage for political
elites and other memory actors to re-perform and conceptualize the past and define their political
agendas within that frame. We identified seven commemorations that were relevant because they
either attracted the country’s political leadership and were of national significance, or were
particularly controversial and therefore provoked debates that would reveal how various actors framed
the nation through rival “truths” about the past. Five commemorations are related to the Second World
War: Bleiburg, Brezovica, Jazovka, Jasenovac and Srb. The two other commemorations represent both
the victim and victor narratives of the Croatian War of Independence (referred to as the Homeland
War, or Domovinski rat, in Croatia): Vukovar and Knin. During the course of 3 years (2014-2017) the
members of the FramNat research team gathered audio-visual evidence of the public speeches
delivered at the 7 commemorations. The speeches were transcribed, tokenized, morphosyntactically
tagged, lemmatized and syntactically parsed using the Reldi service2 and then published as a
searchable corpus. By studying the data provided by the analyses of the corpus, we offer conclusions
about the conceptual representation and cultural distribution of collective (national) identities in
discourse and in official narratives of the past. The main questions that are addressed include the
following: what salient concepts and mappings are applied by the speakers to frame conceptualizations
of the nation and national identity? What are the differences in framing the nation and national identity
between different political actors, institutions and options? Who are the individuals, organizations and
institutions that produce dominant models of representing cultural memory, how and why? One of the
objectives of this chapter is to provide insights into methods and processes that were involved in
gathering and analyzing information on the memory agents and their role in commemorative speeches.
The second section of the chapter discusses the methods and resources used for the gathering and
storing the data, while the third section demonstrates some of the analytic uses of the data for reseach
in the conceptual framing methods.
Gathering the data about the commemorations
The project FRAMNAT gathered empirical data about seven commemoration practices connected to
the wars of the 20th century: Jasenovac, Bleiburg, Brezovica and Jazovka, Srb, Knin and Vukovar
using various methodologies involving audio-visual field recordings, transcription of commemoration
speeches and creation of a text corpus.
Socio-cultural data about the speakers in the commemorations
Most of the commemorations involve a public speech delivered by the speaker, who acts as a memory
agent. The speaker’s role is to conceptualize the historical traumatic event by captivating the attention
and raising the motivation of the listeners, while at the same time providing reasoning patterns and
establishing the culturally normative values3. The conceptualization is performed by the speech
delivered by a speaker and addressed primarily to the assembled audience at the commemoration site,
and secondarily to the wider national audience through the media coverage. Most of the speakers are
connected with some institution or organization, such as the Croatian Parliament, President or
Government, as well as other social organizations such as Catholic, Orthodox church, Anti-fascist
organizations, veteran organizations etc., that partake in the political agenda of the commemoration.
The network of the sixty-four speakers who delivered speeches at seven commemoration sites from
2014-2016. is represented in Illustration 1.
Illustration 1: Network representation of the speakers at each commemoration. The size of the nodes is
represented relative to the amount of connections with other nodes (degree).
The layout of the graph is produced by connecting a speaker to the commemoration site where the
speech was delivered. The majority of the speakers have delivered speeches at only one
commemoration site, but some of them, mostly high ranking political officials, have appeared at
several commemorations, such as the former president Ivo Josipović, who delivered speeches in Knin,
Brezovica, and Jasenovac, as did former Prime Minister Zoran Milanović. Kolinda Grabar-Kitarović,
elected president in January 2015, appeared as a speaker in Knin and Brezovica. Cardinal Josip
Bozanić and other members of the Catholic Church also appeared at several commemorations
including Knin, Vukovar, and Bleiburg. The following three networks represent the presence of the
speakers on a yearly basis from the year 2014 (illustration 2), 2015 (illustration 3), to 2016 (illustration
4).
Illustration 2: Network representation of the speaker’s attendance at commemorations during the year
2014.
Illustration 3: Network representation of the speaker’s attendance at commemorations during the year
2015.
Illustration 4: Network representation of the speaker’s attendance at commemorations during the year
2016.
Audio-visual material
The primary method of the data gathering was related to the fieldwork observation and audio-visual
recordings of the commemorations. Every speech and related commemoration ritual have been audio-
visually recorded by researchers of the FRAMNAT project. The recordings were edited and stored on
a local hard drive as well as published on the FRAMNAT Youtube channel4.
Illustration 5. Screenshot of the FRAMNAT YouTube channel.
The audio-video material allows for the multimodal research of commemoration speeches and related
rituals such as wreath laying ceremonies, clerical rituals and processions. The published material is
organized and tagged according to the year and the commemoration.
Transcriptions of the speeches and corpus creation
The audio-visual data was the basis for the transcription of the texts. The transcription has been
performed by researchers and collaborators on the FRAMNAT project: Iva Davorija, Renato
Stanković and Mirna Gurdon. In order to allow for consistent text analysis, the texts have been stored
as .txt files along with their unique metadata filename: the name of the speaker, date of the speech and
the name of the commemoration. For instance, the speech delivered by Bozanić in Knin in the year
2014 is stored as “bozanic_kn_2014.txt” file. The texts of this FRAMNAT corpus are stored locally in
the Neo4j property database and published on the SketchEngine cloud service5.
FramNat corpus on SketchEngine
The SketchEngine service is used for the text storage, tokenization of text, morphosyntactic tagging,
lemmatization and parsing of the texts using the IHJJ sketch grammar for Croatian6. The FRAMNAT
corpus is published on the SketchEngine service at the following address:
https://the.sketchengine.co.uk/auth/corpus/140810/search.
Tagged FramNat corpus stored in a Neo4j graph database
A locally stored FramNat Corpus for additional ontological analysis of the texts is created using a
custom-developed software application for storing linguistically tokenized, lemmatized and
syntactically parsed digital texts of the Croatian language using the Reldi application service7, py2neo
Python library8 and graph property database Neo4j9. The application uses a pipeline of several
automated processes that comprise: 1) ingesting the texts as data, 2) tokenizing, lemmatizing and
parsing the texts, and 3) storing multiple texts and tokenized, lemmatized and parsed morphosyntactic
information in the graph database. The application pipeline is presented in Illustration 6.
Illustration 6. The pipeline for creating tokenized, lemmatized and syntactically parsed corpus using
the Reldi Api, Neo4j graph database and Py2Neo application.
In the first phase, texts are processed in the form of .txt raw data and sent to the Reldi application
programming interface (https://github.com/clarinsi/reldi-lib-doc). In the second step, the Reldi service
parses the text data and sends back the lexical, morphosyntactic and syntactic values about tokens,
lemmas and dependency structures in the form of a JSON data file. Next, the Python py2neo
application converts the JSON data values stored on a local drive to a Neo4j database using a custom-
made linguistic schema model. The model extracts the JSON key: value structure to represent each
text as an entity with the properties, such as the name of the text, creation date, link to the resource,
etc. The schema of the model stores the linguistic structures of tokens, words, lemmas as entities while
the grammatical and syntactical relations are stored as connections between those entities. Each text is
therefore decomposed into corresponding tokens, words and lemmas according to the described
linguistic schema (Illustration 7).
Data integration
The data about the audio-visual resources, texts and metadata are published on the Google Sheet:
https://docs.google.com/spreadsheets/d/1rXV9x9-Jdpw84nmcOTEJBHnd-S5nu7-
YDYk8zj06sN8/edit?usp=sharing.
This datasheet represents the structure of the information for each commemoration speech, including
the name of the commemoration, description, available information about the commemoration on
Wikipedia, available information about the historical reference on the commemoration topic, place of
the commemoration, GPS coordinates of the commemoration site, year of the commemoration,
speaker name, speaker party affiliation, Wikipedia resources on the speaker, additional resources on
the speaker from media, picture, gender, function, affiliated institution, institution information, text
name, text, estimated number of attendees, video link, and order of speech in the commemoration
event. The various data points are integrated using an ontological model that embeds the
commemoration metadata, and textual and morphosyntactic data with the resources for media research
on the social context. The enriched information is stored in the graph property database Neo4j and
connected via the ontological model shown in Illustration 7. The graph type of database is used to
represent connected data. The graph representation of the commemorations communicates the
structure of connectedness between data points at different phenomenological levels and enables the
quantitative and qualitative exploration of the correlational structure and their construals.
Illustration 7. Ontological model and Database Schema of Commemorative Speech Analysis
The nodes represent the structurally different categories of entities and their directional relations to
component categories. For instance, a text has sentences, and syntactic constructions are part of
sentences. Word forms are part of lemmas, while lemmas are part of texts. Furthermore, each node and
relations have some properties that are expressed in the key: value format. For instance, the Person
category has properties name and gender. Person is connected with the category Commemoration with
a relation ATTEND that has properties: name, place, date, geolocation, picture and video. The
ontologically different levels of the digital phenomena of commemorations are stored in a single
database that enables complex queries within a certain category or between categories. The directed
graph property structure supports the complex analysis of the social identity of actors, their
interactions and institutional affiliations, and the cultural models they express in their speeches.
Analysis of the data
The analytic system comprises of queries that are formulated using the Cypher native programming
language for Neo4j database10 and Py2Neo Python library. The results can be represented as tabular
sets of data or in a network. The basic use of the analytic tools is related to various types of
summarization and statistics. For instance, the thirty most frequent lemmas (basic word forms) or
entity concepts in the FramNat corpora 2014-2016 can be identified with the following query:
MATCH p=(t:Texts)-[r:HAS_lemma]->(l:Lemmas)
WHERE l.lempos ENDS WITH “-n”
WITH l.lempos as Lemma,r.lemmaCountInFile as count
RETURN lema, sum(count) as Sum
ORDER BY Sum DESC
LIMIT 30
The output of this query is represented in a Table 1.
Table 1. Thirty most frequent noun lexical concepts in the FRAMNAT 2014-2016 corpus.
Lemma (lexical concept) Translation Sum
Hrvatska Croatia 486
narod People 323
godina Year 322
čovjek Man 308
žrtva Victim 269
dan Day 226
rat War 219
život Life 195
država State 188
istina Truth 182
mjesto Place 174
zločin Crime 157
sloboda Freedom 154
borba Struggle 150
domovina Homeland 144
branitelj Defender 137
put Path 135
zlo Evil 122
grad City 121
povijest History 121
mir Peace 102
zemlja Land 100
republika Republic 99
ime Name 97
Hrvat Croat 97
riječ Word 95
kraj End 93
Vukovar Vukovar 90
vrijeme Time 89
fašizam Fascism 87
Furthermore, if we want to extract similarities between two speakers on the level of the conceptual use
we can formulate a query that looks for the overlapping lexical concepts (lemmas). The following
Cypher query displays these differences between texts produced by Ivica Glavota and Dražimir Jukić
MATCH p=(s1:Speakers{name:"Zoran Milanović"})-[d1:DELIVERED_SPEECH]->(t1:Texts)-
[l1:HAS_lemma]->(lema:Lemmas)<-[l2:HAS_lemma]-(t2:Texts)<-[d2:DELIVERED_SPEECH]-
(s2:Speakers{name:"Želimir Puljić"})
WHERE lema.lempos ENDS WITH "-n"
WITH distinct(lema.lempos) AS lemm
RETURN lemm
The query returns following nouns: mjesto “place”; naš “our”; budućnost “future“; branitelj
“defender“; izraz “expression“; vjera “faith“; zajedništvo “community“; dijete “child“; dobro “good“;
osjećaj “feeling“; vlast “government“; oko “eye“; srce “heart“; republika “republic“; Hrvatska
“Croatia“; predsjednik “president“; rat “war“; gospodin “mr.“; dan “day“; sloboda “freedom“;
poštovanje “respect“; čovjek “man“; godina “year“; vojska “army“; Europa “Europe“; svijet “world“;
građanin “citizen“; zemlja “land“; dio “part“; obitelj “family“; mir “peace“; život “life“; država
“state“; ime “name“; ponos “pride“; put “path“; prostor “space“; otpor “resistance“; strana “side“;
kraj “end“; prijatelj “friend“; događaj “event“; Hrvat “Croat“; vrijeme “time“; tisuća “thousand“;
želja “wish“; narod “people“; mjesec “moon“; čelo “forehead“; trenutak “moment“; sila “force“;
kultura “culture“; početak “start“; ruka “hand“; riječ “word“; sestra “sister“; zajednica “community“;
poziv “call“; pad “fall“; način “means“; svećenik “priest“; dom “home“; temelj “ground“; čast
“honour“; sredstvo “means“; Radić “Radić“; pozdrav “salute“; inozemstvo “foreign country“; tijelo
“body“; rodbina “family“.
Due to the similarity between lexical concepts expressed by all speakers we can calculate the
proximity of different speakers and represent it by means of the speaker community graph. This
community identification method can be used for discerning the Cultural Model of conceptualization
related to a particular Speaker (Illustration 8).
Illustration 8: The graph of relationships between the 3370 noun lemmas expressed by the 64 speakers.
The size of the labels corresponds to the overall frequency of the nouns connected with the speaker.
The Louvain algorithm for detecting communities11 was applied to the network represented in
Illustration 8 and distinguishes ten communities of speakers. The communities, clustered according to
the similarity of the nouns they used in their speeches, are shown in Table 2.
Table 2. Ten communities of the speakers clustered according to the similarity of the nouns used in
their speeches.
Community Speakers % of the network
activation
1 Josip Bozanić, Franjo Komarica, Vjekoslav Huzjak, Nikola
Kekić, Juraj Jezerinac, student 6, student 5, student 7,
student 9, student 4
17,6 %
2 Franjo Habulin, Milorad Pupovac, Stjepan Mesić, Zoran
Pusić, Milan Surla, Milan Tankosić, Ivanka Roksandić,
Dragan Markovina
18,9 %
3 Mate Uzinić, Želimir Puljić, Đuro Hranić, Ivica Jagodić,
Manda Patko, student 3, student 1, student 2
12,6 %
4 Bruna Esih, Dragan Čović, Ante Kutleša, Željko Reiner,
Idriz ef. Bešić, Zlatko Ževrnja, Borjana Krišto, Aziz ef.
Hasanović, Orest Wilczynski, student 8
11,7 %
5 Kolinda Grabar-Kitarović, Ivo Josipović, Josip Leko, Milan
Bandić, Dražimir Jukić, Ivan Vukić, Dubravka Jurlina
Alibegović, Margareta Mađerić, Madij Ismailov
11,3 %
6 Kristina Ikić Baniček, Nikola Budija 5, 1 %
7 Zoran Milanović, Tomislav Sopta, Maciej Szymanski,
Branko Lustig
8,2 %
8 Boris Prebeg, Frano Čirko 3,7 %
9 Ivica Glavota, Josipa Rimac 3,4 %
10 Nataša Mataušić, Nevenka Marinković, Imam Admir
Muhić, Ivica Vukelić
7,5 %
Lastly, we can perform similar types of queries by connecting the Speaker entities with the affiliated
Institutions. In this way we get the representation of the influence of the institutions in the overall
corpus (Illustration 9)
Illustration 9. The graph of the relationships between the 3370 noun Lemmas expressed by the
representatives of 31 Institutions. The size of the labels and nodes corresponds to the overall frequency
of the nouns connected with a particular Institution. The graph is projected in 3 ordinates: x, y and z.
The projected z ordinate, perceived as the height of the institution nodes, corresponds to the number of
different words connected in the graph.
The graph representation can be interactively explored at the FRAMNAT web site. It is important to
note the connectedness and the structure between the institutions and concepts, represented by the
Force Layout with z ordinate in Illustration 9. The nouns commonly used by many representatives of
the institutions are located in the oval centre of the graph due to the many connections with different
representatives, while the nouns specific and unique to a certain institution extend to the margins.
The data integration allows for the refinement of the various dimensions of the exploration within the
structure of the dataset. For instance, the previous graph (illustration 9) states that the Catholic Church
produced most of the concepts in the dataset. By using the graph structure we can query where were
these speeches performed?
Illustration 10: Network representation of the speaker’s attendance affiliated with the Catholic Church
in Croatia.
The insight of these two graphs is that the institution Catholic Church was active in the Jazovka and
Bleiburg commemorations of the Second World War atrocities, as well as in the Vukovar and Knin
commemorations of the Croatian War of Independence.
Another type of the conceptual analysis is related to the construal of the relevant entities. For instance,
the “dobj” type of syntactic-semantic construction reveals what the object is of some process. So, we
can formulate a query that reveals what is the conceptualization of some entities when construed as a
direct object. In other words, we ask, what you can do with some entity? Here is the query for the
noun domovina or “homeland” construed as a direct object:
MATCH (a:Lemmas{lempos:"domovina-n"})-[r:HAS_dependency{function:"dobj"}]-(b:Lemmas)
WHERE b.lempos ENDS WITH "-v"
WITH max(r.countDep) AS freq, a.lempos AS directObject, b.lempos AS process, r
RETURN process, directObject, r.function AS dependencyType, freq ORDER BY freq DESC
The result of the query is represented in Table 3.
Table 3. The processes that conceptualize the noun domovina or “homeland” as a direct object:
Process directObject dependencyType Freq
imati “have” domovina “homeland” dobj 4
voljeti “love” domovina “homeland” dobj 3
obilježiti “mark” domovina “homeland” dobj 1
štititi “protect” domovina “homeland” dobj 1
uzeti “take” domovina “homeland” dobj 1
voditi “lead” domovina “homeland” dobj 1
graditi “build” domovina “homeland” dobj 1
biti “be” domovina “homeland” dobj 1
poštovati “respect” domovina “homeland” dobj 1
The problem of construal of the abstract social concept HOMELAND is related to the metonymic and
metaphoric conceptual analysis of the cognitive processes12. In this case, the underlying metaphorical
mappings are produced with the construal of HOMELAND as AN OBJECT THAT SOCIAL AGENT
POSSESSES; A THING THAT PSYCHOLOGICAL AGENT LOVES; A VALUABLE GOOD TO
BE PROTECTED|RESPECTED, AN OBJECT THAT CAN BE TAKEN, AN OBJECT THAT CAN
BE LED, A THING THAT CAN BE BUILT.
Conclusion
This chapter deals with the analysis of commemoration practices from the perspective of the
digitization and data integration of various phenomenological levels of analysis. The main goal of this
interdisciplinary research is to understand the process of construing the culturally distributed cognition
and conceptualizations by means of public communication acts and speeches. We have identified the
speakers as the social agents in promoting immediate conceptual and gradual cultural dissemination.
The content of their message is framed by the salient concepts from a cultural model, or worldview,
that speakers share by institutional affiliation. The message is analyzed as text consisting of tokens,
words and lemmas, applying NLP methods to the Croatian language. The value of this multilevel
ontological model is in fostering an interdisciplinary approach through contextualization of data and
targeted usage of digital resources. The contextualization of the data enables holistic insight into the
dynamics of cultural memory practices and their political implications for contemporary culture. The
usage of the digital methods allows for a fine grained quantitative analysis that, due to the graph
property organization of the model, continues to be qualitatively expressive, flexible and non-
reductive.
Bibliography
Ljubešić, Nikola; Klubička, Filip; Agić, Željko; Jazbec, Ivo-Pavao. „New Inflectional Lexicons and
Training Corpora for Improved Morphosyntactic Annotation of Croatian and Serbian” In: Proceedings
of the Tenth International Conference on Language Resources and Evaluation (LREC 2016). Portorož,
Slovenija: European Language Resources Association (ELRA), 4264-4270, 2016.
Pavlaković, Vjeran, and Perak Benedikt. "How Does This Monument Make You Feel? Measuring
Emotional Responses to War Memorials in Croatia.", in: The Twentieth Century in European Memory:
Transcultural Mediation and Reception, eds. Plewa, Barbara Törnquist, and Tea Sindbæk Andersen,
Brill, http://www.brill.com/products/book/twentieth-century-european-memory-2017.
www.cost.eu/module/download/62629, 2017.
Štrkalj Despot, K., Brdar, M., Essert, M., Tonković, M., Perak, B., Ostroški Anić, A., Nahod, B.,
Pandžić, I. MetaNet.HR – Croatian Metaphor Repository. http://ihjj.hr/metafore/, 2015.
Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, Etienne Lefebvre „Fast unfolding of
communities in large networks“, Journal of Statistical Mechanics: Theory and Experiment, 10, 2008.
1 The FRAMNAT project is a four-year project, financed by the Croatian Science Foundation under the number
HRZZ-3782. This paper was supported by this project. 2 Ljubešić, Nikola; Klubička, Filip; Agić, Željko; Jazbec, Ivo-Pavao. „New Inflectional Lexicons and Training
Corpora for Improved Morphosyntactic Annotation of Croatian and Serbian // Proceedings of the Tenth
International Conference on Language Resources and Evaluation (LREC 2016)”. Portorož, Slovenija: European
Language Resources Association (ELRA), 4264-4270, 2016. 3 Pavlaković, Vjeran, and Perak Benedikt. "How Does This Monument Make You Feel? Measuring Emotional
Responses to War Memorials in Croatia.", in: The Twentieth Century in European Memory: Transcultural
Mediation and Reception, eds. Plewa, Barbara Törnquist, and Tea Sindbæk Andersen, Brill,
http://www.brill.com/products/book/twentieth-century-european-memory-2017. 4 FRAMNAT Youtube channel: https://www.youtube.com/channel/UCjarsad6jWsKPi4Z7CdK5Wg. 5 The SketchEngine cloud service: https://the.sketchengine.co.uk. 6 The IHJJ sketch grammar for Croatian: https://the.sketchengine.co.uk/auth/sketch_grammar/356/view/. 7 The Croatian language using Reldi application service: https://github.com/clarinsi/reldi-lib-doc. Ljubešić et al.
2016. 8 Py2neo Python library: http://py2neo.org/v3/. 9 Graph property database Neo4j: https://neo4j.com/. 10 Cypher programming language: https://www.opencypher.org/. 11 Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, Etienne Lefebvre „Fast unfolding of
communities in large networks“, Journal of Statistical Mechanics: Theory and Experiment 2008 (10), 1000. 12 Štrkalj Despot, K., Brdar, M., Essert, M., Tonković, M., Perak, B., Ostroški Anić, A., Nahod, B., Pandžić, I.
(2015). MetaNet.HR – Croatian Metaphor Repository. [Database]. http://ihjj.hr/metafore/.
